Integrating machine learning and statistical testing for studying emotional influence in online review

Purpose This study aims to develop a robust method for analyzing emotional influence in online reviews, addressing limitations in existing approaches. By integrating machine learning with statistical testing, the research seeks to (1) combine the quantitative precision of statistical methods with the qualitative accuracy of machine learning, (2) overcome biases associated with manual voting systems in measuring perceived usefulness and (3) capture the multi-emotion contagion consumers experience during decision-making. The purpose is to provide a more accurate understanding of how different emotional valences (positive and negative) influence perceived usefulness and consumer decisions, offering practical tools for managing brand reputation. Design/methodology/approach This study proposes a hybrid approach combining machine learning and statistical testing to analyze emotional influence in online reviews. First, a dataset of 70,013 reviews for experiential (toothpaste) and search-based (medical apps) products was collected and preprocessed. Perceived usefulness was classified using binary machine learning combined with information entropy. Features were extracted using XGBoost, and sentiment analysis was performed with a fine-tuned BERT model. The Wilcoxon test was applied to assess the significance of emotional influence by comparing sentiment valences (positive/negative) with perceived usefulness. This integrated framework offers a robust and non-parametric analysis of emotional impacts on decision-making. Findings The study reveals that emotional influence varies by product type. For experiential products (e.g. toothpaste), positive sentiment significantly impacts perceived usefulness, while negative sentiment does not. In contrast, for search-based products (e.g. medical apps), both positive and negative sentiments significantly influence perceived usefulness, with negative sentiment having a stronger impact. The proposed hybrid method, combining machine learning and the Wilcoxon test, effectively captures these differences, overcoming limitations of traditional approaches. These findings highlight the importance of tailoring sentiment management strategies based on product characteristics and suggest prioritizing emotional content with higher perceived usefulness for strategic decision-making. Originality/value This study introduces a novel hybrid approach combining machine learning with statistical testing to analyze emotional influence in online reviews, addressing key limitations of traditional methods. By integrating sentiment classification (BERT) with perceived usefulness (information entropy) and significance testing (Wilcoxon test), the study offers a more precise framework for understanding multi-emotion contagion. Unlike prior studies focusing on single-emotion impacts, this research explores the complex interplay of positive and negative emotions across different product types. The findings provide valuable theoretical insights and practical tools for businesses to strategically manage emotional content, optimize consumer decision-making and enhance brand reputation management.